Performance optimization of topic modeling algorithms using a graphic processing unit

Abstract

Text mining can be effectively deployed to improve our understanding of the real world by extracting relevant features such as cultural context from text data. Techniques such as topic models are shown to be useful in automatically extracting the topical or semantic content from unstructured data. Such a system should consume a large amount of text and extract meaningful patterns usually within a specified amount of time. Graphics Processing Unit is increasingly being used for computationally intensive tasks because of the inexpensive, high-performance raw processing power it has to offer. In this paper, we implement, test, and compare various topic modeling algorithms in a Graphics Processing Unit to achieve faster computing time compared to traditional implementations in a Central Processing Unit. The goal is to execute parallel Graphics Processing Unit versions of algorithms, such as Latent Dirichlet Allocation and Latent Semantic Analysis, and quantitatively assess the performance of each algorithm in comparison with serial or multi-core versions of the same topic modeling algorithms. The study aims to provide a comprehensive understanding of the effectiveness of a spectrum of Topic Model algorithms, the merits of such models in the Graphics Processing Unit, and the magnitude of efficiency improvement that can be achieved. Experimental results show that Topic Modeling algorithms can achieve 10x to 40x speedup in the GPU framework.